Today’s news runs along four lines: the economics of AI tighten in opposite directions (DeepSeek cuts prices, Anthropic posts margin, neoclouds raise capacity), Anthropic stages a coordinated lead-up to the Mythos 1 release, the developer-plumbing layer hits a maturity milestone with MCP’s biggest revision since launch, and Washington steps back from binding AI rules after a single phone call.
Economics: Price Wars, Profitability, and the Neocloud Bet
Three stories about where the money is moving: an open-weight permanent price cut, a frontier lab’s margin proof, and an infrastructure thesis with railroad-scale capex behind it.
- DeepSeek makes its 75% V4 Pro discount permanent — What was supposed to be an end-of-month promotion is now the standing price. DeepSeek’s flagship undercuts GPT-5, Claude Opus 4.7, and Gemini 3.5 Flash, widening the cost gap on reasoning-heavy enterprise workloads to roughly 7-to-1 against Western flagships.
- Anthropic’s path to profit runs through falling compute costs — Compute cost per revenue dollar fell from 71 cents in Q1 to 56 cents in Q2, with Claude Code alone at $2.5 billion in revenue and a $559 million profit projection ahead of the October IPO.
- SpaceX became a cloud company and the math is staggering — SpaceX’s Anthropic deal is worth roughly 80% of its 23 years of rocket revenue. The $7.5 trillion buildout forecast runs at 5% of US GDP, on par with the 1880s railroad boom, which is also the template for catastrophic overbuild.
Anthropic Stages Its Next Release
Three coordinated signals point at the same thing: Mythos 1 general availability is close, the marketing is being staged through capability disclosure, and the memory architecture is being reworked to match.
- Anthropic finds Claude Mythos can build complete working exploits — New benchmarks show Mythos Preview turning V8 vulnerabilities into arbitrary code execution, outperforming every other tested model on research Anthropic helped design. Restricted access positioned as the responsible alternative.
- Mythos 1 surfaces in cloud security tools ahead of broader release — Testing Catalog spotted Mythos signatures in Google Cloud and AWS vulnerability discovery programs. The closed-consortium framing has not been updated to match the deployment footprint, and the broader release is now visibly staged.
- Anthropic plans Memory Files: notes sharded across structured documents — Claude’s memory system is being refactored from one growing conversational blob into multiple structured documents organized by topic, project, or context.
Agent Plumbing Grows Up
MCP’s biggest revision since launch, a population-level eval workflow from OpenAI, and a DeepSeek-native coding agent built around prefix-cache economics. The infrastructure layer underneath agents is moving from improvised to specified.
- MCP’s biggest spec revision since launch enters release candidate — Stateless core, formal extensions framework, OAuth/OIDC-aligned authorization, deprecation policy. Final spec ships 2026-07-28 and contains breaking changes.
- OpenAI publishes a macro-eval workflow for agentic systems — The unit of analysis shifts from a single failed trace to patterns across thousands. Lower-level evals grade individual runs; macro-evals cluster the recurring failures the per-run review never sees.
- Reasonix bets on prefix-cache stability for low-cost coding agents — A DeepSeek-native terminal coding agent that optimizes for the cost shape of long-running sessions rather than UX features. Pairs with DeepSeek’s 75% permanent cut for order-of-magnitude lower per-hour cost.
What AI Can’t Do, and What Washington Won’t
Two stories about the limits: agents can index your data faster than humans but cannot tell you what matters; the U.S. government has just stepped back from imposing binding AI rules.
- AI agents can index your qualitative data. They cannot tell you what matters. — Six structured experiments by Shreya Shankar find the same pattern: agents paraphrase rather than analyze, invent new labels for every passage, and produce confident summaries without editorial taste.
- David Sacks talks Trump out of a sweeping AI executive order — An 11th-hour phone call from the VC and current administration AI advisor convinced Trump that mandatory AI regulation would hand the lead to China. The U.S. now has no binding federal framework on the 12-24 month horizon.
Today’s Quick Hits
- ChatGPT now fills out forms from a photo — Upload a form image, describe the details to include, and ChatGPT returns a completed version. The natural-language fill step is the substantive change, not the OCR.
- Perplexity open-sources Bumblebee, a security scanner for developer machines — A read-only scanner that flags risky npm packages, browser extensions, and AI tool configurations on developer workstations.
- Gemini 3.5 Flash (Low) generates 45% fewer tokens than Medium — Developer testing finds the cheapest reasoning tier outperforms the most expensive on software-engineering tasks. If the pattern holds on your workloads, swapping the default tier cuts inference cost roughly in half.